The field of philosophy has long grappled with the complexities of intelligence and understanding, seeking to frame these abstract concepts within an evolving world. The exploration of Large Language Models (LLMs), such as ChatGPT, has fuelled this discourse further. Research by Murray Shanahan contributes to these debates by offering a precise critique of the prevalent terminology and assumptions surrounding LLMs. The language associated with LLMs, loaded with anthropomorphic phrases like ‘understanding,’ ‘believing,’ or ‘thinking,’ forms the focal point of Shanahan’s argument. This terminological landscape, Shanahan suggests, requires a complete overhaul to pave the way for accurate perceptions and interpretations of LLMs.
The discursive journey Shanahan undertakes is enriched by a robust understanding of LLMs, the intricacies of their functioning, and the fallacies in their anthropomorphization. Shanahan advocates for an understanding of LLMs that transcends the realms of next-token prediction and pattern recognition. The lens through which LLMs are viewed must be readjusted, he proposes, to discern the essence of their functionalities. By establishing the disparity between the illusion of intelligence and the computational reality, Shanahan elucidates a significant avenue for future philosophical discourse. This perspective necessitates a reorientation in how we approach LLMs, a shift that could potentially redefine the dialogue on artificial intelligence and the philosophy of futures studies.
The Misrepresentation of Intelligence
The core contention of Shanahan’s work lies in the depiction of intelligence within the context of LLMs. Human intelligence, as he asserts, is characterized by dynamic cognitive processes that extend beyond mechanistic pattern recognition or probabilistic forecasting. The anthropomorphic lens, Shanahan insists, skews the comprehension of LLMs’ capacities, leading to an inflated perception of their abilities and knowledge. ChatGPT’s workings, as presented in the study, offer a raw representation of a computational tool, devoid of any form of consciousness or comprehension. The model generates text based on patterns and statistical correlations, divorced from a human-like understanding of the context or content.
Shanahan’s discourse builds upon the established facts about the inner workings of LLMs, such as their lack of world knowledge, context beyond the input they receive, or a concept of self. He offers a fresh perspective on this technical reality, directly challenging the inflated interpretations that gloss over these fundamental limitations. The model, as Shanahan emphasizes, can generate convincingly human-like responses without possessing any comprehension or consciousness. It is the intricate layering of the model’s tokens, intricately mapped to its probabilistic configurations, that crafts the illusion of intelligence. Shanahan’s analysis breaks this illusion, underscoring the necessity of accurate terminology and conceptions in representing the capabilities of LLMs.
Prediction, Pattern Completion, and Fine-Tuning
Shanahan introduces a paradoxical element of LLMs in their predictive prowess, an attribute that can foster a deceptive impression of intelligence. He breaks down the model’s ability to make probabilistic guesses about what text should come next, based on vast volumes of internet text data. These guesses, accurate and contextually appropriate at times, can appear as instances of understanding, leading to a fallacious anthropomorphization. In truth, this prowess is a statistical phenomenon, the product of a complex algorithmic process. It does not spring from comprehension but is a manifestation of an intricate, deterministic mechanism. Shanahan’s examination highlights this essential understanding, reminding us that the model, despite its sophisticated textual outputs, remains fundamentally a reactive tool. The model’s predictive success cannot be equated with human-like intelligence or consciousness. It mirrors human thought processes only superficially, lacking the self-awareness, context, and purpose integral to human cognition.
Shanahan elaborates on two significant facets of the LLM: pattern completion and fine-tuning. Pattern completion emerges as the mechanism by which the model generates its predictions. Encoded patterns, derived from pre-training on an extensive corpus of text, facilitate the generation of contextually coherent outputs from partial inputs. This mechanistic proficiency, however, is devoid of meaningful comprehension or foresight. The second element, fine-tuning, serves to specialize the LLM towards specific tasks, refining its output based on narrower data sets and criteria. Importantly, fine-tuning does not introduce new fundamental abilities to the LLM or fundamentally alter its comprehension-free nature. It merely fine-tunes its pattern recognition and generation to a specific domain, reinforcing its role as a tool rather than an intelligent agent. Shanahan’s analysis of these facets helps underline the ontological divide between human cognition and LLM functionality.
Revisiting Anthropomorphism in AI and the Broader Philosophical Discourse
Anthropomorphism in the context of AI is a pivotal theme of Shanahan’s work, re-emphasizing its historical and continued role in creating misleading expectations about the nature and capabilities of machines like LLMs. He offers a cogent reminder that LLMs, despite impressive demonstrations, remain fundamentally different from human cognition. They lack the autonomous, self-conscious, understanding-embedded nature of human thought. Shanahan does not mince words, cautioning against conflating LLMs’ ability to mimic human-like responses with genuine understanding or foresight. The hazard lies in the confusion that such anthropomorphic language may cause, leading to misguided expectations and, potentially, to ill-conceived policy or ethical decisions in the realm of AI. This concern underscores the need for clear communication and informed understanding about the true nature of AI’s capabilities, a matter of crucial importance to philosophers of future studies.
Shanahan’s work forms a compelling addition to the broader philosophical discourse concerning the nature and future of AI. It underscores the vital need for nuanced understanding when engaging with these emergent technologies, particularly in relation to their portrayal and consequent public perception. His emphasis on the distinctness of LLMs from human cognition, and the potential hazards posed by anthropomorphic language, resonates with philosophical arguments calling for precise language and clear delineation of machine and human cognition. Furthermore, Shanahan’s deep dive into the operation of LLMs, specifically the mechanisms of pattern completion and fine-tuning, provides a rich contribution to ongoing discussions about the inner workings of AI. The relevance of these insights extends beyond AI itself to encompass ethical, societal, and policy considerations, a matter of intense interest in the field of futures studies. Thus, this work further strengthens the bridge between the technicalities of AI development and the philosophical inquiries that govern its application and integration into society.
Abstract
Thanks to rapid progress in artificial intelligence, we have entered an era when technology and philosophy intersect in interesting ways. Sitting squarely at the centre of this intersection are large language models (LLMs). The more adept LLMs become at mimicking human language, the more vulnerable we become to anthropomorphism, to seeing the systems in which they are embedded as more human-like than they really are. This trend is amplified by the natural tendency to use philosophically loaded terms, such as “knows”, “believes”, and “thinks”, when describing these systems. To mitigate this trend, this paper advocates the practice of repeatedly stepping back to remind ourselves of how LLMs, and the systems of which they form a part, actually work. The hope is that increased scientific precision will encourage more philosophical nuance in the discourse around artificial intelligence, both within the field and in the public sphere.
Talking About Large Language Models